autolearn

CRITICAL: The `backend/sandbox.py` explicitly bypasses the sandboxing mechanism (`logger.warning(f"Executing skill {skill_name} directly (no sandbox) due to pickling limitations")`), meaning dynamically generated Python code is executed directly in the main server process using `exec()`. This allows for arbitrary remote code execution via API endpoints (`/skills/generate`, `/skills/register`, `/mcp` tools/call), which is an extremely severe vulnerability. Any unauthenticated access to these endpoints would allow complete system compromise. This is the most critical security flaw, making the system unsafe to run in its current state without a robust sandbox or process isolation.